Hi Tomasz, Thanks for submitting the CURAND wrapper for inclusion in PyCUDA!
On Thu, 03 Feb 2011 12:01:55 +0100, Tomasz Rybak <[email protected]> wrote: > Sorry if you have received my previous mail (from 2011-02-01 22:52) > but I have not got any reply - so here I am sending it again. > > I have seen your email from 2010-12-26. I was in Germany, so I was > able to send reply on 2011-01-11. I have fixed code and removed > repetitions. CURAND wrapper already gained some interest in PyCUDA > mailing list - Martin Laprise used it and pointed that it can be > optimised in mail from 2011-01-18. > > I am attaching patch (diff to master > f2ae507a357a3bf7949beff03963995b655fa547) I've put (a fair bit!) of work into your wrapper. You can find the results of my work here: http://git.tiker.net/pycuda.git/shortlog/refs/heads/curand-wrapper-v2-from-tomasz (i.e. branch curand-wrapper-v2-from-tomasz of the main repo) I'm generally happy with the state of the code, but a few issues remain that I would like you to sort out before I merge to master: - All computations are submitted as just a single block (grid=(1,1)) That cannot be a good use of the GPU. We should default to creating at least (2-3)x(number of SMs) blocks of generators, so we can fill up the GPU. - The Sobol' direction vectors need to come from a very specific set to make sense, see curandGetDirectionVectors32 in the CURAND docs. We should probably call/wrap this function to get those vectors. Further, each generator should use a different vector, rather than the same one. - The Sobol' initialization needs to be worked out. In particular, I would like both generators to do something sensible if they're initialized without arguments. Andreas
pgpII0CM923ik.pgp
Description: PGP signature
_______________________________________________ PyCUDA mailing list [email protected] http://lists.tiker.net/listinfo/pycuda
